BrainSeg-Net: Brain Tumor MR Image Segmentation via Enhanced Encoder-Decoder Network.

Area of Science:

• Medical Imaging and Artificial Intelligence
• Neuro-oncology and Computational Pathology

Background:

• Accurate segmentation of Magnetic Resonance (MR) brain tumor images is critical for diagnosis and treatment planning.
• Existing neural network models struggle with segmenting small tumor regions due to loss of spatial information in deeper layers and class imbalance.

Purpose of the Study:

• To propose an advanced encoder-decoder model, BrainSeg-Net, for efficient and accurate brain tumor segmentation.
• To enhance the identification of small-scale tumor sub-regions and address challenges of class imbalance in MR images.

Main Methods:

• Developed BrainSeg-Net, an encoder-decoder architecture incorporating a Feature Enhancer (FE) block to aggregate multi-level features.
• Implemented a custom loss function to manage class imbalance issues.
• Evaluated performance on three benchmark datasets: BraTS2017, BraTS2018, and BraTS2019, segmenting Enhancing Core (EC), Whole Tumor (WT), and Tumor Core (TC).

Main Results:

• BrainSeg-Net demonstrated significant improvements in segmenting brain tumor sub-regions (EC, WT, TC) compared to baseline and state-of-the-art methods.
• The FE block effectively preserved and utilized location and spatial information, crucial for identifying small tumor areas.
• The custom loss function successfully mitigated challenges posed by imbalanced tumor class distribution.

Conclusions:

• BrainSeg-Net offers a superior approach to MR brain tumor segmentation by effectively leveraging enhanced location and spatial features.
• The model's ability to handle small-scale regions and class imbalance makes it a valuable tool for clinical diagnosis.
• This architecture represents a significant advancement over existing brain MR image segmentation techniques.